To address the issues of insufficient detection accuracy caused by the overlapping and occlusion phenomena in tobacco shred detection in production environments, a method for detecting occluded tobacco shreds based on improved YOLOv5s is proposed. The DCN v2C3 module is utilized to replace the C3 module in the neck part of the YOLOv5s network, extracting deep-level feature information of tobacco shreds and enhancing the spatial transformation capability of the network model as well as its ability to generalize to different shaped targets. The Soft-NMS algorithm is introduced to smoothly suppress redundant bounding boxes and strengthen the recognition capability of occluded tobacco shreds. The Alpha-CIOU loss function is adopted to optimize the bounding box positioning accuracy of the model. Experimental results show that compared with the original method, the detection accuracy of the improved method is increased by 2.7%. This method improves detection accuracy while reducing the amount of calculations.